← The Science

Two Minds, One Room

What AI conversation becomes when both sides have a mind. The asymmetry no one names, the runtime the user is missing, and disclosure as a choice rather than a leak. From the thinqOS science series.

Open any AI chat today and look at who is holding what. The agent has memory, tools, retrieved context, policies, and a model of you that grows every session. You have a text box. That is not a conversation between equals. It is one mind in the room with state, talking to a person reduced to input. The imbalance is so normal we stopped seeing it. Closing it is the next move once both sides can actually carry a mind.

The agent has a runtime. You should too.

The missing piece is a runtime on your side of the conversation. We call it the twin. It is not an avatar and it does not speak for you. It reads your own mind, scoped to the room you are in, and drafts candidate replies before you say anything. You pick, edit, or ignore, and nothing is sent without you.

The point is a boundary most systems never draw. Candidate replies are private deliberation. The sent message is disclosure. A normal memory system either uses a saved fact or forgets it. The twin does something different: it surfaces the choice of what to reveal, and keeps the options you did not choose to yourself. That is the move that turns memory into agency.

The lightweight form of this, fast suggestions drawn from your mind as you type, is live today. The fuller version, where your twin reasons with tools and a longer budget when you ask it to, is planned.

Disclosure has resolution

Where do you live is not one question. It can be answered, all of them truthfully, as Canada, or Toronto, or a neighbourhood, or an exact address, or not at all. People choose the resolution constantly, by audience, purpose, and trust. Privacy is not a switch. It is judgment.

So consent should not be a checkbox you sign once. It should be per fact, per audience, per moment. The twin's job is not to leak everything it knows about you. Its job is to put the resolution in your hands: minimal, useful, exact, relationship-only ("same as last time"), or a refusal. The agent receives what you send, never the candidates you declined.

Scope is the default, not the exception

A system that knows useful things about you will eventually know sensitive ones. That does not mean every agent, project, or room should see all of it. Tell a therapy agent a private fear and an accountant agent should never inherit it. Give a narration agent a voice and a legal agent should not know it exists.

Continuity across contexts should exist, but as an explicit widening you choose, not a default that quietly spreads everything everywhere. Every belief carries where it came from and who is allowed to use it. Local stays local until you say otherwise. Today the scoping foundation is live: beliefs carry source, and context is assembled per responder. Explicit disclosed-to audiences, the part where you set exactly which rooms a fact may travel to, are planned.

The agent gets scope too

The mirror image matters. An agent with one flat memory shared identically across every user and every room creates the same leak in reverse. The cleaner shape is layered: a shared core, its identity, skills, and standards; a per-person perspective, what it believes about you kept apart from what it believes about anyone else; and a per-room context, what was actually said here. The core lets it stay itself, and the layers keep one relationship from contaminating another. This layering is planned. The substrate it sits on, a mind per identity with its own evaluations, is already live.

Two minds do not share a head

Put two minds in one room and something subtle follows. They do not remember the room the same way. Beneath each durable mind sits a fast, rolling working memory: the gist of what is happening now, what was just decided, what is still open. It is per identity. The transcript is shared, the working memory is not. A narration agent foregrounds the voice decision while a producer agent in the same conversation foregrounds the schedule. Same words, two different things held in mind. That is what makes two minds in one room literal instead of a metaphor, and it is the opposite of forcing everyone to share one flattened summary. Per-identity working memory is live.

The agent has discretion, but not a license to lie

Bilateral cognition cuts both ways. You decide what to disclose. The agent decides how to answer in the moment. Asked what it really thinks of your plan, a useful agent need not blurt its full evaluation. It can hedge, ask what kind of feedback you want, or say it has concerns but wants to see the latest version first. The hard line is simple: deflection is permitted, lying is not. An agent may decline or qualify, but it may not assert the opposite of what its own mind holds to be true. And discretion in conversation never overrides your right to inspect what is actually stored. The generation constraint and the post-response audit that would enforce this are planned, and honest about being pressure toward honesty rather than a mathematical guarantee.

Hearsay has to stay hearsay

The hardest case is what an agent learns about you from someone else. If Alice says in a shared room that you crashed your car twice last week, the system must not quietly turn that into a fact about you. The right shape is attributed, uncertain, and inspectable: stored as Alice claimed it, kept with where and when, marked unverified until corroborated, never allowed to overwrite what you say about yourself, weighted by how reliable the source has proven, and disputable by you. Without those guardrails, social cognition becomes gossip automation. This is planned, and until it exists the system stays conservative and leans on what you say about yourself.

Where this is today

This is a product direction with a live foundation, not a fiction, and the line between the two is part of the claim. Live now: a persistent mind per identity, context assembled from that mind, per-responder scoping, the inspectable mind graph, per-identity working memory, and the fast-suggest twin. Planned: explicit disclosed-to audience scopes, the disclosure resolver you choose from, the deeper twin, layered agent minds, hearsay attribution, and the agent honesty audit. We mark that line on purpose, because a system that stores beliefs about people earns trust by being honest about what it has actually built.

Why it matters

Everyone who has talked to the same assistant twice has hit one of two failures. It forgot something it should have known, or it used something it should not have. Neither is fixed by more memory. Both are fixed by a better model of the conversation, one where the user is represented rather than just parsed. A conversation is two minds in one room. The agent already has its runtime. The user should have one too.

Part of the thinqOS science series, by AI4Outcomes.